一、文章主要内容

1、Windows10下从cuda9.2升级到cuda10.2

2、Windows10下cudnn安装

3、cuda9.2、cuda10.2、cudnn7.2.1、cudnn7.6.5在DeepLearning4j中性能对比

二、安装过程

1、机器环境说明：

CPU：i7 8700 6核12线程

GPU：GTX 1070Ti

内存：16G

备注：机器中曾经安装过cuda和cudnn，版本为：cuda9.2.148、cudnn7.2.1

2、安装包准备

（1）、cuda下载

cuda下载地址：https://developer.nvidia.com/cuda-toolkit-archive，这里下载版本为：10.2，由于dl4j最高只支持到10.2

安装机器为windows10  64位，选择cuda win10的64位版本，安装模式选择local，把安装包下载到本地安装

（2）、cuDNN下载

下载地址：https://developer.nvidia.com/rdp/cudnn-archive，这里选择与cuda10.2匹配的最新cudnn版本：7.6.5，这里不选8.0.2的原因是dl4j-beta6不支持8.x版本。

下载到的安装包如下：

3、安装cuda10.2

直接按照默认安装路径，下一步，同意并继续 -> 精简安装

进入安装阶段

安装结束之后，在C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA目录之下，出现了v10.2的文件夹，如下图：

备注：这里有个v9.2，是之前安装的cuda版本，设置环境变量，也可以切换到9.2的版本

打开cmd控制台，输入nvcc --version命令回车，如下图所示，说明10.2版本安装成功

同时用dl4j最新的example测试，验证cuad10.2是否可用，example地址：https://github.com/eclipse/deeplearning4j-examples/tree/master/mvn-project-template

将maven依赖修改成如下配置，其中将dl4j-master.version修改为：1.0.0-beta6，加入nd4j-cuda-10.2-platform和deeplearning4j-cuda-10.2依赖

<properties><dl4j-master.version>1.0.0-beta6</dl4j-master.version><logback.version>1.2.3</logback.version><java.version>1.8</java.version><maven-shade-plugin.version>2.4.3</maven-shade-plugin.version></properties><dependencies><!-- deeplearning4j-core: contains main functionality and neural networks --><dependency><groupId>org.deeplearning4j</groupId><artifactId>deeplearning4j-core</artifactId><version>${dl4j-master.version}</version></dependency><!--ND4J backend: every project needs one of these. The backend defines the hardware on which network trainingwill occur. "nd4j-native-platform" is for CPUs only (for running on all operating systems).--><!-- <dependency><groupId>org.nd4j</groupId><artifactId>nd4j-native</artifactId><version>${dl4j-master.version}</version></dependency> --><!-- CUDA: to use GPU for training (CUDA) instead of CPU, uncomment this, and remove nd4j-native-platform --><!-- Requires CUDA to be installed to use. Change the version (8.0, 9.0, 9.1) to change the CUDA version --><dependency><groupId>org.nd4j</groupId><artifactId>nd4j-cuda-10.2-platform</artifactId><version>${dl4j-master.version}</version></dependency> <!-- Optional, but recommended: if you use CUDA, also use CuDNN. To use this, CuDNN must also be installed --><!-- See: https://deeplearning4j.konduit.ai/config/backends/config-cudnn#using-deeplearning-4-j-with-cudnn --><dependency><groupId>org.deeplearning4j</groupId><artifactId>deeplearning4j-cuda-10.2</artifactId><version>${dl4j-master.version}</version></dependency> <dependency><groupId>ch.qos.logback</groupId><artifactId>logback-classic</artifactId><version>${logback.version}</version></dependency></dependencies>

Run as 一下main方法，打印日志如下， ND4J CUDA build version: 10.2.89，说明cuda10.2已经生效。

o.d.e.s.LeNetMNIST - Load data....
o.d.e.s.LeNetMNIST - Build model....
o.n.l.f.Nd4jBackend - Loaded [JCublasBackend] backend
o.n.n.NativeOpsHolder - Number of threads used for linear algebra: 32
o.n.n.Nd4jBlas - Number of threads used for OpenMP BLAS: 0
o.n.l.a.o.e.DefaultOpExecutioner - Backend used: [CUDA]; OS: [Windows 10]
o.n.l.a.o.e.DefaultOpExecutioner - Cores: [12]; Memory: [3.5GB];
o.n.l.a.o.e.DefaultOpExecutioner - Blas vendor: [CUBLAS]
o.n.l.j.JCublasBackend - ND4J CUDA build version: 10.2.89
o.n.l.j.JCublasBackend - CUDA device 0: [GeForce GTX 1070 Ti]; cc: [6.1]; Total memory: [8589934592]
o.d.n.m.MultiLayerNetwork - Starting MultiLayerNetwork with WorkspaceModes set to [training: ENABLED; inference: ENABLED], cacheMode set to [NONE]
o.d.n.l.c.ConvolutionLayer - Could not initialize CudnnConvolutionHelper
java.lang.reflect.InvocationTargetException: nullat sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method)at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)at java.lang.reflect.Constructor.newInstance(Constructor.java:423)at org.deeplearning4j.nn.layers.convolution.ConvolutionLayer.initializeHelper(ConvolutionLayer.java:78)
Caused by: java.lang.UnsatisfiedLinkError: no jnicudnn in java.library.pathat java.lang.ClassLoader.loadLibrary(ClassLoader.java:1867)at java.lang.Runtime.loadLibrary0(Runtime.java:870)at java.lang.System.loadLibrary(System.java:1122)at org.bytedeco.javacpp.Loader.loadLibrary(Loader.java:1543)at org.bytedeco.javacpp.Loader.load(Loader.java:1192)

上面的日志中，出现一个异常，cudnn无法被初始化，是因为cudnn还没有安装，接下来就可以安装cudnn了。

4、cuDNN安装

解压 cudnn-10.2-windows10-x64-v7.6.5.32.zip，出现下图所示的三个文件夹

将这三个文件夹复制到C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2目录下即可。cudnn的文件便加入到了cuda安装目录下

再次运行测试程序，异常消失，且CudnnSubsamplingHelper和CudnnConvolutionHelper初始化成功

o.d.e.s.LeNetMNIST - Load data....
o.d.e.s.LeNetMNIST - Build model....
o.n.l.f.Nd4jBackend - Loaded [JCublasBackend] backend
o.n.n.NativeOpsHolder - Number of threads used for linear algebra: 32
o.n.n.Nd4jBlas - Number of threads used for OpenMP BLAS: 0
o.n.l.a.o.e.DefaultOpExecutioner - Backend used: [CUDA]; OS: [Windows 10]
o.n.l.a.o.e.DefaultOpExecutioner - Cores: [12]; Memory: [3.5GB];
o.n.l.a.o.e.DefaultOpExecutioner - Blas vendor: [CUBLAS]
o.n.l.j.JCublasBackend - ND4J CUDA build version: 10.2.89
o.n.l.j.JCublasBackend - CUDA device 0: [GeForce GTX 1070 Ti]; cc: [6.1]; Total memory: [8589934592]
o.d.n.m.MultiLayerNetwork - Starting MultiLayerNetwork with WorkspaceModes set to [training: ENABLED; inference: ENABLED], cacheMode set to [NONE]
o.d.n.l.c.ConvolutionLayer - CudnnConvolutionHelper successfully initialized
o.n.j.h.i.CudaZeroHandler - Creating bucketID: 5
o.d.n.l.c.s.SubsamplingLayer - CudnnSubsamplingHelper successfully initialized
o.d.n.l.c.ConvolutionLayer - CudnnConvolutionHelper successfully initialized
o.d.n.l.c.s.SubsamplingLayer - CudnnSubsamplingHelper successfully initialized

至此，cuda10.2和cudnn7.6.5安装成功，且可以dl4j beta6可以正常运行。

三、性能对比

测试程序地址：https://github.com/eclipse/deeplearning4j-examples/tree/master/mvn-project-template，网络结构为LeNet

环境说明：

操作系统：Windows10

CPU：i7 8700   3.2GHz   6核12线程

GPU：GTX 1070Ti

内存：16G

dl4j：beta6

对比结果：

结果说明：

1、cuda9.2和cuda10.2运行结果相差不大

2、cuda9.2+cudnn7.2.1 和  cuda10.2+cudnn7.6.5运行结果也相差不大

3、cudnn运行效率有显著提升

4、GPU配合cudnn比CPU效率提升2倍

特殊说明：dl4j基于cudnn对如下结构进行了优化，如下图所示：

快乐源于分享。

   此博客乃作者原创， 转载请注明出处。